Neural BER Estimation for Low-Power ECC Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current error correction coding techniques in memory devices and wireless baseband circuitry face challenges with increasing area and power needs, leading to higher costs and longer development times, and introduce errors due to bit flips in non-volatile memory devices, which can degrade storage and transmission efficiency.
Innovation Solution
The use of multi-layer neural networks and recurrent neural networks to estimate the bit error rate (BER) of encoded data, allowing for efficient decoding by comparing the estimated BER to a threshold and reducing errors before decoding, thereby improving processing speed and reducing computational and power resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If complex error correction coding techniques are used, then error correction capability is improved, but area and power needs increase
Solution Approach 1:
The patent replaces traditional mechanical/combinatorial error correction decoding systems with a neural network-based system. The neural network learns error patterns and performs decoding through trained weights and activations, substituting the conventional step-by-step mechanical decoding process with a data-driven computational model that achieves better error correction with reduced hardware footprint.
Solution Approach 2:
The patent changes the operational parameters of the error correction system by using learnable weights and activation functions instead of fixed decoding algorithms. The neural network adapts its parameters through training on error patterns, allowing it to optimize error correction performance while using fewer resources compared to traditional fixed-parameter error correction codes.
2Reliability
If complex error correction coding techniques are used, then error correction capability is improved, but power consumption increases
Solution Approach 1:
The patent replaces power-intensive traditional error correction decoding circuits with a neural network implementation that can be optimized for low power consumption. The neural network's structure allows for efficient computation patterns and can be deployed on low-power hardware platforms, reducing overall power consumption while maintaining or improving error correction capability.
Solution Approach 2:
The patent employs a neural network model that can be trained once and then deployed for repeated use with minimal additional computational overhead. The trained weights and parameters serve as a compact, low-power representation of error correction logic that can be quickly applied to multiple data blocks without requiring repeated complex computations.
3Productivity
If traditional decoding methods are used, then decoding can be performed, but processing speed is reduced
Solution Approach 1:
The patent performs preliminary training of the neural network on extensive error patterns and data before actual decoding operations. This pre-computation phase allows the network to learn optimal decoding strategies, which are then rapidly applied during actual operation. The preliminary action of training enables faster real-time decoding while maintaining high accuracy.
Solution Approach 2:
The patent substitutes traditional sequential decoding algorithms with a neural network that can process multiple error patterns in parallel through its distributed structure. This substitution enables faster decoding speeds while the network's learned representations maintain or improve decoding accuracy compared to conventional methods.
Data Source
AI summary
Examples described herein utilize multi-layer neural networks, such as multi-layer recurrent neural networks, to estimate a bit error rate (BER) of encoded data based on a retrieved version of encoded data (e.g., data encoded using one or more encoding techniques) from a memory. The neural networks may have nonlinear mapping and distributed processing capabilities which may be advantageous to estimate a BER of encoded data, e.g., to facilitate decoding of the encoded data. In this manner, neural networks described herein may be used to improve or facilitate aspects of decoding at ECC decoders, e.g., by comparing an estimated BER to a threshold (e.g., a threshold BER level) prior to decoding of the encoded data. For example, an additional NN activation indication may be provided, e.g., to indicate that the encoded data may be decoded or to indicate that error present in the encoded data is to be reduced.